Generation of map layer is extremely helpful for visual ease in navigation applications. Google Maps provide many layers for the convenience of their user base. Out of them, the most used layers are satellite imagery and default map layer. In this work, we attempt to generate automatic map layer from optical satellite images. We enforce a conditional GAN with a patch discriminator for generating realistic map layer with crisp higher frequency details.
We also selected a L1 loss function to address the lower frequency details in the output. This work is inspired by the pix2pix model. [Paper]
All the models are developed in keras and can be easily extended to other domains.
The dataset used for this experiment can be found here
Here is the model's performance after 109600 epochs.
In this test sample, the map is generated aesthetically and accurately as per satellite map. The ground truth is observed to have false information.
In this test sample, the vegetation area's map is not generated perfectly by the model.
